Deep learning based system identification of industrial integrated grinding circuits

Abstract

Energy efficiency and maximum productivity in ore beneficiation processes can be ensured when integrated grinding circuits function in an optimal fashion. The complexity of first principles based models prevents online implementation of control and optimization algorithms, thus, creating the need for the development of accurate data-based models. In this work, deep recurrent neural networks (DRNNs) are implemented for nonlinear system identification of 3 input 6 output integrated grinding circuit from an industrial lead-zinc ore beneficiation set-up. Optimal long short term memory networks (LSTMs) with maximum predictability are obtained by solving a novel multi-objective framework for DRNN architecture design. The optimal LSTMs are trained and validated on pseudo random binary sequence (PRBS) signal with an accuracy of 99%, and tested successfully on unseen random Gaussian sequence (RGS) signal. Comprehensive comparison with conventional tools for nonlinear system identification, such as wavelet networks, is performed to show the efficacy of proposed optimal LSTMs.